Fourier tensor-base illumination compensation for face recognition

Illumination variation remains one of the key challenges in real face recognition systems. This paper proposes a novel Fourier tensor-based illumination compensation (FTIC) algorithm to enhance the image at the preprocessing step to ensure the facial recognition rate. First, the color face image is separated into R, G, and B color channels and then mapped to the Fourier domain to extract the magnitudes of the three color channels containing most lighting information. Second, a tensor in which Fourier's magnitudes of three color channels are tensor's frontal slices is constructed and then decomposed by higher-order singular value decomposition to get more detail components. Third, the core tensor is linearly changed by multiplying with our designed compensation coefficients using a Gaussian template as reference. Fourth, the new three color channels’ magnitudes are constructed based on the latest core tensor and our designed inverse factor. Finally, the compensated color face image is built using an inverse Fourier transform. The experiments conducted on CMU-PIE reveal that the proposed algorithm yields clearer, smoother, and more natural face images and improves the accuracy of real face recognition systems.